Landslide Monitoring using Convolutional Autoencoders

Landslides are one of the major emergency situations that lead to huge damage and human victims. A monitoring system to reduce the effects of landslides is mandatory under these circumstances. In this paper, we propose an automatic landslide detection algorithm based on the use of multispectral remote sensing data. The compressed representation of the spatial and spectral data is obtained using the encoder part of a Convolutional Autoencoder architecture, whereas the separation of the landslide area is achieved through a K-Means unsupervised clustering algorithm. The data used for experiments was collected for areas in Romania severely affected by landslides, for a period of approximately 2-3 months before and after the event. Satellite images can be acquired from the Earth Explorer platform and Google Earth Engine.